Tess-4.5-27B, a project by Migtissera distributed in GGUF format, flips one of the most watched metrics in local Large Language Model deployment. It does not present itself as a bigger or better-trained model: its defining trait is producing fewer tokens for the same result. According to data released by the author, already known for Synthia, this Qwen fine-tune generates on average 25% fewer tokens than its base across the same 150 BenchLocal scenarios, reasoning included. The reduction is not uniform: on CLI-40 tasks it reaches 32% with a higher score, on data extraction it hits 43%, while on bug finding it stops at 21%. The cut concentrates where the work is hardest, rather than applying a generic trim.

For teams running self-hosted models, a token is not just a unit of text. It is compute load, latency, progressive occupancy of the context window and, ultimately, TCO. Producing fewer tokens to reach the same result means faster responses, longer sessions before the context window fills and less energy per inference. In an on-premises context, where GPU capacity is a physical constraint and you cannot simply scale by adding cloud instances, the same hardware can serve more requests or slightly larger models without upgrading the fleet.

The effect compounds when the model is paired with a draft head: every token saved reduces verification steps and further accelerates generation. This shifts incentives for fine-tuning developers: the goal is no longer just the highest score, but the best result per token produced. That direction rewards lean models and moves competition from training at ever-larger scale to inference optimization.

The winners are teams with local GPUs and constrained hardware budgets, runtime vendors that can demonstrate lower latency, and organizations that must respect data sovereignty without moving workloads to the cloud. The losers, at least in part, are vendors that rely on ever-larger models as their only sales argument: token efficiency introduces an evaluation criterion that cannot be bought with more VRAM.

The numbers should be read with due caution: these are the author's internal benchmarks on BenchLocal scenarios, not an independent verification. The GGUF format, however, makes the model immediately usable on local stacks and lends itself to quick testing by those who want to verify results on their own infrastructure.